What AI does to a state depends on where it is aimed

Governments are being asked to adopt AI quickly, often on old systems and stretched budgets, and the officials doing it are judged later on consequences that were hard to see when they decided.

In 2024 HMRC piloted a check that matched child benefit claims against Home Office travel records. Experienced staff ran it, with an upfront check of employment records before any enquiry opened, and 54% of the cases investigated proved ineligible. For the August 2025 rollout, HMRC used temporary staff and, judging the employment check too complicated for them, moved it to the end of the process. The National Audit Office found the change was made without considering its wider implications. Payments were suspended as soon as an enquiry opened, for 23,794 claimants. An estimated 38% proved ineligible, and savings came to £60m against an initial assumption of £128m. HMRC paused new enquiries in October 2025 and restarted in March 2026 with the employment check back up front and new cases capped at 50 a week.

Each decision was defensible on its own. Together they produced a live system without the pilot's safeguard, more cases than staff could check, and savings counted before they arrived. The check was data matching, not AI, but AI produces the same pattern faster and at larger scale.

Two simulations

I built two simulations of AI inside government. The first covers the UK, Kenya and the Democratic Republic of the Congo (DRC) over 15 years; the second, Kenya over 30 years, with its economy and public finances. Each runs thousands of versions of the future, varying what is unknown, such as how often AI errs. In each, a plan is compared with the same government not using AI itself.

Harm means wrongful official decisions, weighted by how long each lasts. A scandal is harm public enough to force a pause on AI in public services. A plan succeeds if services improve and trust holds, without more force against protest or enforcement outgrowing its checks. Today's approach is my estimate of how a government uses AI now, since none publishes it; for the UK, about 40% on services and 37% on security and enforcement. These are teaching models, not forecasts.

Four choices decided the outcome

What AI is pointed at. On today's approach, the UK simulation had services improve in only 46% of futures, trust fall in 86%, and a scandal in 94%. Enforcement errors are the costly kind: a wrong letter can be put right, but a wrongful sanction or stopped payment comes back as appeals and repeat contact, uses up the time AI saved, and grows into a scandal. The UK plan that did best aimed AI mainly at front-line services, acted only on what staff could check, reinvested most gains in services and put safeguards in law. It succeeded in 74% of futures, against 8% for today's approach, with scandals in 38% and harm cut to under a third. Aiming at the front line mattered most. Neither simulation counts what enforcement AI recovers, so the lesson is about risk: add enforcement uses last, once checking is in place.

How much of its output is acted on. On today's UK approach, 50% more review staff cut harm by under a tenth. Acting only on as much AI output as staff could check cut it by about a fifth. Testing each use before it scales, and dropping uses that fail, roughly halved it. Pilots can look better than live use, as HMRC's did. Set the volume you act on by what staff can check, as HMRC's capped restart does, and judge each use on its live results.

Whether safeguards bind. In the Kenya simulation, safeguards with no AI at all (impact assessments, keeping service records out of security use, audit logs, limits on the use of force) raised growth slightly in about 19 futures in 20, even after I charged them for slower adoption, oversight bodies and compliance work for firms. People share data and use digital services more when they trust how data are handled. Combined with AI, safeguards delivered more than the sum of their parts. The largest single uncertainty was how far records built for services get turned to police and security use in a crisis, and the better AI gets, the more useful those records become, so the separation matters more over time. In the UK simulation, the best plan failed most often where AI was most capable, and in nearly nine of ten such failures, enforcement power had outgrown the institutions that check it. Safeguards held only in guidance weakened when a later government changed course; in law they held. That is partly my assumption: the UK's Data (Use and Access) Act 2025, which eased the rules on automated decisions, shows law can loosen too.

What comes first. In the Kenya simulation I tested every order of three phases: accountability (oversight and redress), services, and tax. Accountability and services first, with broad tax collection last, did best at every horizon. In the first ten years the best order delivered nearly three times the services gain of the worst, which put tax first. A push for revenue raised about half the revenue of the plan that waited. In the simulation, better services build the trust that makes people pay; pressing many small taxpayers before that trust exists drains it. By 30 years every order with safeguards ended in roughly the same place. If results are needed within one term, build accountability and services first and collect broadly later.

What differs by country

The plan built for the UK succeeded in 74% of UK futures, 42% of Kenyan ones and 19% of DRC ones. In the UK, the main constraint is counting savings before they arrive: every 1% of services given up through staff cuts freed about £0.5bn a year. Reporting the savings actually achieved, as the National Audit Office recommended in July 2026, keeps that trade a decision for ministers. In Kenya and the DRC, testing each use before it scales mattered far more, because AI errs more often on thinner records, the assumption I am least sure of. In the DRC, where charities, churches and donors deliver much of what people receive, an aid cut roughly halved services in both simulations.

Kenya

Kenya exemplifies how AI could impact a state that has a strong technology scene in a low-middle income economy. In the Kenya simulation, a stress test with about half of AI on security and broad tax collection (not a reading of Kenya's strategy) produced a growth gain in under 25% of futures by 2041, against about 80% for a services-led plan with safeguards. Both simulations find every plan with safeguards or led by services beats today's approach over 15 years. They disagree on whether today's approach leaves services better or worse than no AI, which turns on how often AI errs in Kenyan public services.

Limits

Many settings are my judgment. The ranking of plans holds under testing; the levels do not (the UK plan's 74% ranges from 37% to 93% depending on how success is defined). Whether services rise under today's approach depends on my estimate of it. The safeguard-cost and Kenya-disagreement results come from later extra runs. Neither simulation has been checked against real outcomes, as I lack such data. Treat each result as a question to test.

What I would ask of you

Ultimately, these simulations are only rough ideas and designs, and should be taken as such, as they lack sensitive organized data from these states and assumptions had to be made. Further, the future, what emerges, the true consequences of compounding factors, are often hard to simulate. If you work on AI in government, you can tell me where these simulations are wrong, above all how your AI is split between purposes and how often live tools err. Models and results are available from info@arcede.com on request.

This final version was reviewed and polished by AI. The purpose of this article is to point toward a framework for thinking about these things, not necessarily to establish a definitive plan.